On October 21, at GTC Berlin, Nvidia announced capacity-reservation agreements with six manufacturers: Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron. Jensen Huang's line sums up the moment: “Demand is extraordinary. Capacity is the constraint.”
What Nvidia Actually Announced
According to Nvidia's October 21, 2026 press release, the company is working with six partners to expand manufacturing capacity in the stages that come after the chip is made on the wafer: substrates, interposers, testing and optics. This isn't a story about a new processor; it's about how much it costs, and how long it takes, to deliver the ones that are already selling.
How Each Piece Works (and Why It Matters)
Substrates: the base the chip is mounted on, which connects it to the rest of the system. IBIDEN (Japan) receives a "multibillion-dollar" capacity advance to triple its Ono plant, with additional output from the first quarter of 2028. Unimicron is expanding substrates in Taiwan, with new capacity from the second quarter of 2027 and full capacity by mid-2028. Samsung Electro-Mechanics joins for the first time as a Nvidia substrate supplier, with factories in Sejong and Busan.
Interposers and advanced packaging: Amkor plans to raise its interposer capacity in South Korea by more than 300% by the end of 2028. TechTimes adds that this comes on top of a July 2026 partnership worth US$1.5 billion, focused on Arizona.
Testing: every GPU is verified before it ships. KYEC receives a US$1.1 billion advance for new US facilities. One detail worth keeping straight: in its filing to the market, KYEC described a prepayment of up to NT$45 billion (roughly US$1.4 billion) tied to its US plan, and the contract was signed on October 13. The sources consulted don't explain how the two figures relate, so we treat them as separate numbers.
Co-packaged optics (CPO): SPIL aims to multiply its optical-engine capacity by 15. TrendForce, cited by TechTimes, warns that CPO switches compete for the same 2.5D/3D packaging resources as the AI chips themselves: expanding one piece doesn't solve the other on its own.
The Underlying Number: US$279 Billion in Supply Commitments
TechTimes, citing Nvidia's regulatory filing, reports that supply commitments went from US$119 billion to US$279 billion in a single quarter, and that the gap between supply and demand is expected to persist at least through the fiscal year ending in early 2028. That's consistent with what we've already analyzed about the commitments and guarantees in Nvidia's 10-Q: Nvidia is paying up front to secure capacity that doesn't exist yet.
What This Means for a Company Buying or Using AI
1. Compute pricing and availability are decided downstream. If your plan depends on getting GPUs or GPU instances in 2027, the limiting factor will be packaging, testing and optics, not just the chip design.
2. Negotiate reservations, not just prices. If Nvidia prepays for capacity, large customers do the same through commitment contracts. A mid-sized company in Guatemala or LATAM should compare a capacity reservation with its cloud provider against paying on demand.
3. Design for alternatives. Smaller models, quantization and secondary providers reduce risk if GPU supply stays tight through 2028.
Sources: Nvidia press release (Oct 21, 2026), TechTimes (Oct 22, 2026) and Dealroom on the KYEC filing.